How Far Has AI Actually Come?

Two or three years ago, when many people were first encountering large language models, they learned a phrase that sounded technically sophisticated — AI merely predicts the next word.

The statement is not entirely wrong. During pretraining, a large language model mainly learns by predicting the next token. When it produces an answer, it also normally outputs one token after another. Strictly speaking, a token is not necessarily a complete word. It may be a Chinese character, part of an English word, a punctuation mark, or a fragment of code.

The real problem lies in the word “merely”.

Next-token prediction describes one underlying way in which a model is trained and produces output. It does not tell us what capabilities the model can develop through that training, and it certainly cannot define what a complete AI system — made up of models, tools, memory, and feedback mechanisms — is capable of doing.

It is like saying that a computer is merely a collection of transistors switching between zero and one, or that a piano performance is merely a set of hammers striking strings. As descriptions of underlying mechanisms, both statements are correct. Yet neither explains how a large software system operates or how a Bach fugue comes into being.

Some people still use “predicting the next word” to deny that AI can reason, as if the model were making a random guess among a few nearby words at every step. The reality is already much more complicated. Anthropic’s research into internal model mechanisms found that a model may plan later parts of an expression even though it produces text one word at a time. In one rhyming experiment, the model first formed a choice about the final word and then organised the preceding line around it. In multi-step questions, researchers also observed distinguishable intermediate conceptual processing. This does not prove that AI thinks as humans do, and it certainly does not prove that it is conscious. It does, however, show that “word-by-word output” is not the same as “looking only at the next word”. Anthropic’s research on model mechanisms https://www.anthropic.com/research/tracing-thoughts-language-model

More importantly, the AI systems now in practical use are no longer isolated base language models. They undergo instruction tuning, preference training, and reinforcement learning. They can devote more computation before answering, try different paths, inspect intermediate results, and change approach after detecting an error. They can also search for information, read files, run programs, call databases and external interfaces, and then continue making decisions in response to the results returned by those tools.

Earlier chatbots mainly completed a single question-and-answer exchange. AI agents are now entering a cycle of observation, action, inspection, and correction. They do more than generate a passage of text. They can maintain a goal over a longer process, work within constraints, call tools, inspect the outcome, and decide what to do next. Frontier systems can already work continuously for hours, while some specially designed tasks can continue for days.

Even the finest horse-drawn carriage does not become a car simply by adding more horses. A car is not a faster carriage. It changes the source of power, the method of control, and the wider transport system. If we continue to judge cars by the standards of the carriage, we will only ask whether a car can pull a carriage more effectively. We will fail to see that a different technological form has appeared.

AI is now going through a similar change. It is no longer just a chat window that answers questions more fluently. It is gradually becoming an operational system capable of entering real working environments.

Software development provides the clearest example. AI can now enter large codebases, trace relationships between files and modules, understand call chains, implement features, run tests, locate faults, change code, and verify the results. It is no longer dealing only with a function or a few dozen lines of code. It is working with the interdependent constraints inside a software system.

The claim that “AI can write a million-line software system” is no longer merely promotional exaggeration. In 2026, OpenAI disclosed an internal project that began with an empty repository. Five months later, the project contained roughly one million lines across the application, tests, infrastructure, continuous integration, documentation, and internal tools. About 1,500 pull requests had been merged, and the product already had hundreds of internal users. Every line in the repository had been generated by Codex, with no code written directly by human engineers. OpenAI’s report on the million-line project https://openai.com/index/harness-engineering/

This example must still be understood accurately. It was not a case of an AI receiving a single prompt and producing one million lines of reliable software by itself overnight. Human engineers continued to set goals, break down tasks, design the environment, and establish testing standards and feedback mechanisms. The more accurate conclusion is that, when people provide direction, constraints, and systems of verification, AI can now participate continuously in producing a real software system at the scale of a million lines.

This shows precisely what has changed. Human work is beginning to move away from writing every line of code directly and towards defining goals, designing constraints, building feedback environments, and judging final results. AI is taking on an increasingly large share of concrete execution.

The same change is visible in the social sciences and in philosophical discussion. When concepts and source materials are clearly bounded, AI can compare positions, separate premises from conclusions, maintain stable definitions, identify gaps in an argument, propose counterexamples, and revise a line of reasoning in response to them.

A 2026 preprint asked AI systems to recreate the analytical programs behind 48 social-science papers, using only the papers’ descriptions of their methods and their original data. All 48 papers had already been independently confirmed as reproducible. Among the best-performing agent systems, the reproduced regression coefficients agreed with the direction of the original results more than 85 per cent of the time, and more than 70 per cent fell within the original 95 per cent confidence intervals. This is no longer ordinary text organisation. It involves interpreting research methods, implementing programs, processing data, and comparing results. Social-science reproduction study https://arxiv.org/html/2604.21965v1

Of course, this does not mean that AI reasoning is inherently rigorous, much less that AI has become a source of truth. AI can still misunderstand material, cite sources incorrectly, overlook hidden premises, or invent a superficially complete argument that does not actually hold. Genuine rigour comes from a process that combines stable concepts, source verification, counterexample testing, program execution, and human review. AI can now participate deeply in that process, but it cannot automatically guarantee the outcome.

We also need to distinguish capability, consciousness, and responsibility. The fact that AI can perform complex work does not mean that it possesses human consciousness. Its ability to contribute to judgement does not mean that legal and moral responsibility should be handed over to it. Conversely, the fact that AI cannot bear full human responsibility does not show that it lacks the capacity to analyse, plan, and execute complex tasks. When these questions are mixed together, the discussion easily loses its subject.

What is particularly revealing is that, when some people confront AI, they are no longer mainly interested in what it can actually do. They keep searching instead for what it still cannot do.

It makes mistakes, therefore it is not intelligent. It needs prompts, therefore it has no capability of its own. It lacks human experience, therefore it cannot understand. It cannot bear responsibility, therefore everything it does is merely imitation. Whenever one boundary is crossed, the line marking “real ability” is moved farther away.

This reaction is not difficult to understand. AI is beginning to touch human knowledge, professions, experience, and creativity. Repeatedly emphasising what AI cannot do temporarily preserves an area that appears to belong exclusively to us. It allows us to continue believing that the previous order of capabilities has not fundamentally changed. This is a natural form of psychological self-defence.

The problem is not that people doubt AI. Studying its errors, risks, and limits is essential. The real problem arises when we are willing to see only what it cannot do and refuse to revise our earlier judgement in response to new facts.

Caution and retreat are not the same thing. Caution continually asks under what conditions AI is reliable, where it fails, what supervision it requires, and how the boundary should be redrawn in light of the evidence. Retreat decides in advance that AI cannot possess a certain capability, then treats every failure as proof and explains every success as accident, imitation, or wordplay.

We do not need to worship AI, nor do we need to declare it omnipotent. Today’s AI remains highly uneven. It can sometimes complete expert-level complex tasks and sometimes fail on surprisingly simple ones. It is powerful but unstable. It can reason without guaranteeing correctness, and it can act while still requiring direction and constraint.

Those limitations, however, no longer justify continuing to use a map drawn two or three years ago.

The most reliable way to understand present-day AI is not to keep debating what it must be unable to do in theory. It is to place it inside real, complex, verifiable tasks and observe how much of the work it can actually carry, where it fails, and what people still need to provide. Only then can we see clearly the new relationship taking shape between humans and AI.

Recognising how far AI has already developed is not a surrender to it, nor is it a denial of human value. It simply requires us to respect what is actually happening.

In a rapidly changing world, openness is not blind optimism. It is a basic capacity for recognising reality. What now prevents us from understanding AI may no longer be the complexity of the technology, but our insistence on holding a map from two years ago and refusing to enter the world as it is today.